metadata

license: mit
datasets:
  - yztian/PRIM
language:
  - en
  - de
  - fr
  - cs
  - ro
  - ru
metrics:
  - bleu
  - comet
pipeline_tag: translation

PRIM: Towards Practical In-Image Multilingual Machine Translation (EMNLP 2025 Main)

📄Paper arXiv | 💻Code GitHub

Introduction

This repository provides the VisTrans model, trained as part of our work PRIM: Towards Practical In-Image Multilingual Machine Translation.

The VisTrans model is an end-to-end model for In-Image Machine Translation, which handles the visual text and background information in the image separately, with a two-stage training and multi-task learning strategy.

The model is trained on MTedIIMT, and tested on PRIM (see ./PRIM directory). It is also trained and tested on IIMT30k (see ./IIMT30k directory).

Inference

For inference and detailed usage instructions, please refer to our GitHub repository.

Citation

If you find our work helpful, we would greatly appreciate it if you could cite our paper:

@misc{tian2025primpracticalinimagemultilingual,
      title={PRIM: Towards Practical In-Image Multilingual Machine Translation}, 
      author={Yanzhi Tian and Zeming Liu and Zhengyang Liu and Chong Feng and Xin Li and Heyan Huang and Yuhang Guo},
      year={2025},
      eprint={2509.05146},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2509.05146}, 
}